Efficient Context and Schema Fusion Networks for Multi-Domain Dialogue State Tracking

Page created by Manuel Walsh
 
CONTINUE READING
Efficient Context and Schema Fusion Networks for Multi-Domain Dialogue State Tracking
Efficient Context and Schema Fusion Networks for Multi-Domain
                                                                      Dialogue State Tracking
                                                                    Su Zhu, Jieyu Li, Lu Chen∗ and Kai Yu∗
                                                 State Key Laboratory of Media Convergence Production Technology and Systems
                                                 MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University
                                                          SpeechLab, Department of Computer Science and Engineering
                                                                Shanghai Jiao Tong University, Shanghai, China
                                                        {paul2204,oracion,chenlusz,kai.yu}@sjtu.edu.cn

                                                                 Abstract
                                            Dialogue state tracking (DST) aims at estimat-
                                            ing the current dialogue state given all the pre-
arXiv:2004.03386v4 [cs.CL] 7 Oct 2020

                                            ceding conversation. For multi-domain DST,
                                            the data sparsity problem is a major obstacle
                                            due to increased numbers of state candidates
                                            and dialogue lengths. To encode the dialogue
                                            context efficiently, we utilize the previous di-
                                            alogue state (predicted) and the current dia-
                                            logue utterance as the input for DST. To con-
                                            sider relations among different domain-slots,
                                            the schema graph involving prior knowledge
                                            is exploited. In this paper, a novel context and
                                            schema fusion network is proposed to encode
                                            the dialogue context and schema graph by us-            Figure 1: An example of multi-domain dialogues. Ut-
                                            ing internal and external attention mechanisms.         terances at the left side are from the system agent, and
                                            Experiment results show that our approach can           utterances at the right side are from a user. The dia-
                                            outperform strong baselines, and the previous           logue state of each domain is represented as a set of
                                            state-of-the-art method (SOM-DST) can also              (slot, value) pairs.
                                            be improved by our proposed schema graph.

                                        1   Introduction                                            users across different domains (Budzianowski et al.,
                                                                                                    2018; Eric et al., 2019). As shown in Fig. 1,
                                        Dialogue state tracking (DST) is a key component            the dialogue covers three domains (i.e., Hotel,
                                        in task-oriented dialogue systems which cover cer-          Attraction and Taxi). The goal of multi-
                                        tain narrow domains (e.g., booking hotel and travel         domain DST is to predict the value (including
                                        planning). As a kind of context-aware language              NONE) for each domain-slot pair based on all the
                                        understanding task, DST aims to extract user goals          preceding dialogue utterances. However, due to
                                        or intents hidden in human-machine conversation             increasing numbers of dialogue turns and domain-
                                        and represent them as a compact dialogue state,             slot pairs, the data sparsity problem becomes the
                                        i.e., a set of slots and their corresponding values.        main issue in this field.
                                        For example, as illustrated in Fig. 1, (slot, value)           To tackle the above problem, we emphasize that
                                        pairs like (name, huntingdon marriott hotel) are            DST models should support open-vocabulary based
                                        extracted from the dialogue. It is essential to build       value decoding, encode context efficiently and in-
                                        an accurate DST for dialogue management (Young              corporate domain-slot relations:
                                        et al., 2013), where dialogue state determines the
                                        next machine action and response.                             1. Open-vocabulary DST is essential for real-
                                           Recently, motivated by the tremendous growth                  world applications (Wu et al., 2019; Gao et al.,
                                        of commercial dialogue systems like Apple Siri,                  2019; Ren et al., 2019), since value sets for
                                        Microsoft Cortana, Amazon Alexa, or Google As-                   some slots can be very huge and variable (e.g.,
                                        sistant, multi-domain DST becomes crucial to help                song names).
                                            ∗
                                                The corresponding authors are Lu Chen and Kai Yu.     2. To encode the dialogue context efficiently, we
attempt to get context representation from the      et al., 2014b,a; Yu et al., 2015), or jointly learn lan-
       previous (predicted) dialogue state and the         guage understanding in an end-to-end way (Hender-
       current turn dialogue utterance, while not con-     son et al., 2014b,c). These methods heavily rely on
       catenating all the preceding dialogue utter-        hand-crafted features and complex domain-specific
       ances.                                              lexicons for delexicalization, which are difficult
                                                           to extend to new domains. Recently, most works
    3. To consider relations among domains and             about DST focus on encoding dialogue context
       slots, we introduce the schema graph which          with deep neural networks (such as CNN, RNN,
       contains domain, slot, domain-slot nodes and        LSTM-RNN, etc.) and predicting a value for each
       their relationships. It is a kind of prior knowl-   possible slot (Mrkšić et al., 2017; Xu and Hu, 2018;
       edge and may help alleviate the data imbal-         Zhong et al., 2018; Ren et al., 2018).
       ance problem.
                                                           Multi-domain DST Most traditional state track-
   To this end, we propose a multi-domain dia-             ing approaches focus on a single domain, which
logue state tracker with context and schema fusion         extract value for each slot in the domain (Williams
networks (CSFN-DST). The fusion network is ex-             et al., 2013; Henderson et al., 2014a). They can
ploited to jointly encode the previous dialogue state,     be directly adapted to multi/mixed-domain conver-
the current turn dialogue and the schema graph by          sations by replacing slots in a single domain with
internal and external attention mechanisms. After          domain-slot pairs (i.e. domain-specific slots) (Ra-
multiple layers of attention networks, the final rep-      madan et al., 2018; Gao et al., 2019; Wu et al.,
resentation of each domain-slot node is utilized to        2019; Zhang et al., 2019; Kim et al., 2019). De-
predict the corresponding value, involving context         spite its simplicity, this approach for multi-domain
and schema information. For the value prediction,          DST extracts value for each domain-slot indepen-
a slot gate classifier is applied to decide whether a      dently, which may fail to capture features from slot
domain-slot is mentioned in the conversation, and          co-occurrences. For example, hotels with higher
then an RNN-based value decoder is exploited to            stars are usually more expensive (price range).
generate the corresponding value.                          Predefined ontology-based DST Most of the
   Our proposed CSFN-DST is evaluated on Mul-              previous works assume that a predefined ontology
tiWOZ 2.0 and MultiWOZ 2.1 benchmarks. Abla-               is provided in advance, i.e., all slots and their val-
tion study on each component further reveals that          ues of each domain are known and fixed (Williams,
both context and schema are essential. Contribu-           2012; Henderson et al., 2014a). Predefined
tions in this work are summarized as:                      ontology-based DST can be simplified into a value
                                                           classification task for each slot (Henderson et al.,
     • To alleviate the data sparsity problem and
                                                           2014c; Mrkšić et al., 2017; Zhong et al., 2018; Ren
       enhance the context encoding, we propose
                                                           et al., 2018; Ramadan et al., 2018; Lee et al., 2019).
       exploiting domain-slot relations within the
                                                           It has the advantage of access to the known can-
       schema graph for open-vocabulary DST.
                                                           didate set of each slot, but these approaches may
     • To fully encode the schema graph and dia-           not be applicable in the real scenario. Since a full
       logue context, fusion networks are introduced       ontology is hard to obtain in advance (Xu and Hu,
       with graph-based, internal and external atten-      2018), and the number of possible slot values could
       tion mechanisms.                                    be substantial and variable (e.g., song names), even
                                                           if a full ontology exists (Wu et al., 2019).
     • Experimental results show that our approach         Open-vocabulary DST Without a predefined on-
       surpasses strong baselines, and the previous        tology, some works choose to directly generate
       state-of-the-art method (SOM-DST) can also          or extract values for each slot from the dialogue
       be improved by our proposed schema graph.           context, by using the encoder-decoder architecture
                                                           (Wu et al., 2019) or the pointer network (Gao et al.,
2     Related Work
                                                           2019; Ren et al., 2019; Le et al., 2020). They can
Traditional DST models rely on semantics ex-               improve the scalability and robustness to unseen
tracted by natural language understanding to pre-          slot values, while most of them are not efficient in
dict the current dialogue states (Young et al., 2013;      context encoding since they encode all the previous
Williams et al., 2013; Henderson et al., 2014d; Sun        utterances at each dialogue turn. Notably, a multi-
domain dialogue could involve quite a long history,
e.g., MultiWOZ dataset (Budzianowski et al., 2018)
contains about 13 turns per dialogue on average.
Graph Neural Network Graph Neural Net-
work (GNN) approaches (Scarselli et al., 2009;
Veličković et al., 2018) aggregate information from
graph structure and encode node features, which
can learn to reason and introduce structure infor-
mation. Many GNN variants are proposed and also                 Figure 2: An example of schema graph. Domain nodes
applied in various NLP tasks, such as text clas-                are in orange, slot nodes are in green and domain-slot
sification (Yao et al., 2019), machine translation              nodes are in blue.
(Marcheggiani et al., 2018), dialogue policy opti-
mization (Chen et al., 2018, 2019) etc. We intro-               different domain-slot pairs and exploit them as an
duce graph-based multi-head attention and fusion                additional input to guide the context encoding, we
networks for encoding the schema graph.                         formulate them as a schema graph G = (V, E)
                                                                with node set V and edge set E. Fig. 2 shows an
3    Problem Formulation                                        example of schema graph. In the graph, there are
                                                                three kinds of nodes to denote all domains D, slots
In a multi-domain dialogue state tracking problem,
                                                                S, and domain-slot pairs O, i.e., V = D ∪ S ∪ O.
we assume that there are M domains (e.g. taxi,
                                                                Four types of undirected edges between different
hotel) involved, D = {d1 , d2 , · · · , dM }. Slots
                                                                nodes are exploited to encode prior knowledge:
included in each domain d ∈ D are denoted
as a set S d = {sd1 , sd2 , · · · , sd|S d | }.1 Thus, there        1. (d, d0 ): Any two domain nodes, d ∈ D and
                                                                       d0 ∈ D, are linked to each other.
are J possible domain-slot pairs totally,       PM Odm=
{O1 , O2 , · · · , OJ }, where J =                m=1 |S   |.
                                                                    2. (s, d): We add an edge between slot s ∈ S
Since different domains may contain a same
                                                                       and domain d ∈ D nodes if s ∈ S d , .
slot, we denote all distinct N slots as S =
{s1 , s2 , · · · , sN }, where N ≤ J.                               3. (d, o) and (s, o): If a domain-slot pair o ∈ O
   A dialogue can be formally represented as                           is composed of the domain d ∈ D and slot
{(A1 , U1 , B1 ), (A2 , U2 , B2 ), · · · , (AT , UT , BT )},           s ∈ S, there are two edges from d and s to
where At is what the agent says at the t-th turn, Ut                   this domain-slot node respectively.
is the user utterance at t turn, and Bt denotes the
corresponding dialogue state. At and Ut are word                    4. (s, s0 ): If the candidate values of two different
sequences, while Bt is a set of domain-slot-value                      slots (s ∈ S and s0 ∈ S) would overlap, there
triplets, e.g., (hotel, price range, expensive). Value                 is also an edge between them, e.g., destination
vtj is a word sequence for j-th domain-slot pair                       and departure, leave at and arrive by.
at the t-th turn. The goal of DST is to correctly
predict the value for each domain-slot pair, given              4     Context and Schema Fusion Networks
the dialogue history.                                                 for Multi-domain DST
   Most of the previous works choose to con-                    In this section, we will introduce our approach
catenate all words in the dialogue history,                     for multi-domain DST, which jointly encodes the
[A1 , U1 , A2 , U2 , · · · , At , Ut ], as the input. How-      current dialogue turn (At and Ut ), the previous
ever, this may lead to increased computation time.              dialogue state Bt−1 and the schema graph G by
In this work, we propose to utilize only the cur-               fusion networks. After that, we can obtain context-
rent dialogue turn At , Ut and the previous dialogue            aware and schema-aware node embeddings for all
state Bt−1 to predict the new state Bt . During the             J domain-slot pairs. Finally, a slot-gate classifier
training, we use the ground truth of Bt−1 , while the           and RNN-based value decoder are exploited to ex-
previous predicted dialogue state would be used in              tract the value for each domain-slot pair.
the inference stage.                                               The architecture of CSFN-DST is illustrated in
Schema Graph To consider relations between                      Fig. 3, which consists of input embeddings, con-
   1
     For open-vocabulary DST, possible values for each slot     text schema fusion network and state prediction
s ∈ S d are not known in advance.                               modules.
Figure 3: The overview of the proposed CSFN-DST. It takes the current dialogue utterance, the previous dialogue
state and the schema graph as the input and predicts the current dialogue state. It consists of an embedding layer,
context and schema fusion networks, a slot-gate classifier and an RNN-based value decoder.

4.1   Input Embeddings                                     nodes are arranged as G = d1 ⊕ · · · ⊕ dM ⊕ s1 ⊕
Besides token and position embeddings for encod-           · · ·⊕sN ⊕o1 ⊕· · ·⊕oJ . Each node embedding is ini-
ing literal information, segment embeddings are            tialized by averaging embeddings of tokens in the
also exploited to discriminate different types of          corresponding domain/slot/domain-slot. Positions
input tokens.                                              embeddings are omitted in the graph. The edges
(1) Dialogue Utterance We denote the represen-             of the graph are represented as an adjacency ma-
tation of the dialogue utterances at t-th turn as          trix AG whose items are either one or zero, which
a joint sequence, Xt = [CLS] ⊕ At ⊕; ⊕Ut ⊕                 would be used in the fusion network. To empha-
[SEP], where [CLS] and [SEP] are auxiliary to-             size edges between different types of nodes can be
kens for separation, ⊕ is the operation of sequence        different in the computation, we exploit node types
concatenation. As [CLS] is designed to capture             to get segment embeddings.
the sequence embedding, it has a different segment         4.2   Context and Schema Fusion Network
type with the other tokens. The input embeddings
of Xt are the sum of the token embeddings, the             At this point, we have input representations HG     0 ∈
                                                             |G|×d       Xt      |X  |×d     Bt−1      |B    |×d
segmentation embeddings and the position embed-            R       m , H0 ∈ R      t    m , H0     ∈R    t−1     m,

dings (Vaswani et al., 2017), as shown in Fig. 3.          where |.| gets the token or node number. The con-
(2) Previous Dialogue State As mentioned be-               text and schema fusion network (CSFN) is utilized
fore, a dialogue state is a set of domain-slot-value       to compute hidden states for tokens or nodes in Xt ,
triplets with a mentioned value (not NONE). There-         Bt−1 and G layer by layer. We then apply a stack
fore, we denote the previous dialogue state as             of L context- and schema-aware self-attention lay-
                                                                                                  Xt    Bt−1
Bt−1 = [CLS] ⊕ Rt−1      1               K , where
                              ⊕ · · · ⊕ Rt−1               ers to get final hidden states, HGL , HL , HL     . The
K is the number of triplets in Bt−1 . Each triplet         i-th layer (0 ≤ i < L) can be formulated as:
d-s-v is denoted as a sub-sequence, i.e., R =                      Xt      B                      Xt          Bt−1
                                                           HG               t−1              G
                                                            i+1 , Hi+1 , Hi+1 = CSFNLayeri (Hi , Hi , Hi              )
d ⊕ - ⊕ s ⊕ - ⊕ v. The domain and slot names are
tokenized, e.g., price range is replaced with “price
range”. The value is also represented as a token
sequence. For the special value DONTCARE which             4.2.1 Multi-head Attention
means users do not care the value, it would be re-         Before describing the fusion network, we first in-
placed with “dont care”. The input embeddings of           troduce the multi-head attention (Vaswani et al.,
Bt−1 are the sum of the token, segmentation and            2017) which is a basic module. The multi-head
position embeddings. Positions are re-enumerated           attention can be described as mapping a query and
for different triplets.                                    a set of key-value pairs to an output, where the
(3) Schema Graph As mentioned before, the                  query, keys, values, and output are all vectors. The
schema graph G is comprised of M domain nodes,             output is computed as a weighted sum of the val-
N slot nodes and J domain-slot nodes. These                ues, where the weight assigned to each value is
computed by a compatibility function of the query                     function (Ba et al., 2016). FFN(x) is a feed-
with the corresponding key.                                           forward network (FFN) function with two fully-
   Consider a source sequence of vectors Y =                          connected layer and an ReLU activation in between,
      |Y |
{y i }i=1 where y i ∈ R1×dmodel and Y ∈ R|Y |×dmodel ,                i.e., FFN(x) = max (0, xW1 + b1 ) W2 + b2 .
                                                |Z|
and a target sequence of vectors Z = {z i }i=1                           Similarly, more details about updating
                                                                              Bt−1
where z i ∈ R1×dmodel and Z ∈ R|Z|×dmodel . For                       HX   t
                                                                        i , Hi     are described in Appendix A.
each vector y i , we can compute an attention vector                     The context and schema-aware encoding can
ci over Z by using H heads as follows:
                                                                      also be simply implemented as the original trans-
                   (h)       (h)
           (y i WQ )(z j WK )>                (h)
                                         exp(eij )                    former (Vaswani et al., 2017) with graph-based
 (h)                              (h)
eij =            p             ; aij = P|Z|       (h)                 multi-head attentions.
                   dmodel /H            l=1 exp(eil )
           |Z|
 (h)
           X      (h)      (h)                 (1)        (H)         4.3     State Prediction
ci     =         aij (z j WV ); ci = Concat(ci , · · · , ci     )WO
           j=1                                                        The goal of state prediction is to produce the next
                                                                      dialogue state Bt , which is formulated as two
where 1 ≤ h ≤ H, WO ∈ Rdmodel ×dmodel , and
   (h)   (h)    (h)                                                   stages: 1) We first apply a slot-gate classifier for
WQ , WK , WV ∈ Rdmodel ×(dmodel /H) . We can
                                                                      each domain-slot node. The classifier makes a de-
compute ci for every y i and get a transformed ma-
                                                                      cision among {NONE, DONTCARE, PTR}, where
trix C ∈ R|Y |×dmodel . The entire process is denoted
                                                                      NONE denotes that a domain-slot pair is not men-
as a mapping MultiHeadΘ :
                                                                      tioned at this turn, DONTCARE implies that the
                    C = MultiHeadΘ (Y, Z)                       (1)   user can accept any values for this slot, and PTR
                                                                      represents that the slot should be processed with a
Graph-based Multi-head Attention To apply the                         value. 2) For domain-slot pairs tagged with PTR,
multi-head attention on a graph, the graph adja-                      we further introduced an RNN-based value decoder
cency matrix A ∈ R|Y |×|Z| is involved to mask                        to generate token sequences of their values.
nodes/tokens unrelated, where Aij ∈ {0, 1}. Thus,
 (h)                                                                  4.3.1    Slot-gate Classification
eij is changed as:
                                                                      We utilize the final hidden vector of j-th domain-
                                                                      slot node in G for the slot-gate classification, and
                  (h)        (h) >
           (yi W√Q )(z j WK )
     (h)                            , if Aij = 1                      the probability for the j-th domain-slot pair at the
   eij =             dmodel /H
          
             −∞,                      otherwise                       t-th turn is calculated as:
                                                                                gate
and Eqn. (1) is modified as:                                                   Ptj     = softmax(FFN(HG
                                                                                                      L,M +N +j ))

             C = GraphMultiHeadΘ (Y, Z, A)                      (2)   The loss for slot gate classification is
Eqn. (1), can be treated as a special case of Eqn.                                        T X
                                                                                            J
                                                                                                         gate      gate
                                                                                          X
(2) that the graph is fully connected, i.e., A = 1.                           Lgate = −             log(Ptj     · (ytj )> )
                                                                                          t=1 j=1
4.2.2 Context- and Schema-Aware Encoding
Each layer of CSFN consists of internal and ex-                                 gate
                                                                      where ytj is the one-hot gate label for the j-th
ternal attentions to incorporate different types of                   domain-slot pair at turn t.
inputs. The hidden states of the schema graph G at
the i-the layer are updated as follows:                               4.3.2    RNN-based Value Decoder
                                                                      After the slot-gate classification, there are J 0
       IGG = GraphMultiHeadΘGG (HG    G    G
                                 i , Hi , A )                         domain-slot pairs tagged with PTR class which
                               Xt
     EGX = MultiHeadΘGX (HG
                          i , Hi )
                                                                      indicates the domain-slot should take a real value.
                                                                                                                    gate
                                          Bt−1                        They are denoted as Ct = {j|argmax(Ptj ) =
     EGB = MultiHeadΘGB (HG
                          i , Hi                 )
                                                                      PTR}, and J 0 = |Ct |.
                             G
       CG =       LayerNorm(Hi      + IGG + EGX + EGB )                 We use Gated Recurrent Unit (GRU) (Cho et al.,
     HG
      i+1 = LayerNorm(CG + FFN(CG ))
                                                                      2014) decoder like Wu et al. (2019) and the soft
                                                                      copy mechanism (See et al., 2017) to get the final
where AG is the adjacency matrix of the schema                        output distribution Ptjvalue,k over all candidate to-
graph and LayerNorm(.) is layer normalization                         kens at the k-th step. More details are illustrated in
Appendix B. The loss function for value decoder is           Domain       Slots                    Train Valid Test
                                                            Restaurant    area, food, name,        3813 438 437
             T X X                                                        price range, book day,
             X                                                            book people, book
Lvalue = −                 log(Ptjvalue,k · (ytj
                                              value,k >
                                                     ) )                  time
             t=1 j∈Ct k                                          Hotel    area, internet, name,    3381 416     394
                                                                          parking, price range,
         value,k                                                          stars, type, book day,
where ytj        is the one-hot token label for the j-th
                                                                          book people, book
domain-slot pair at k-th step.                                            stay
   During training process, the above modules can                Train    arrive by, day, de-      3103 484     494
                                                                          parture, destination,
be jointly trained and optimized by the summations                        leave at, book people
of different losses as:                                          Taxi     arrive by, departure,    1654   207   195
                                                                          destination, leave at
                                                            Attraction    area, name, type         2717 401 395
               Ltotal = Lgate + Lvalue
                                                                            Total                  8438 1000 1000
5     Experiment                                                   Table 1: Data statistics of MultiWOZ2.1.
5.1    Datasets
We use MultiWOZ 2.0 (Budzianowski et al., 2018)            (Hashimoto et al., 2017). We do a grid search over
and MultiWOZ 2.1 (Eric et al., 2019) to evaluate           {4, 5, 6, 7, 8} for the layer number of CSFN on the
our approach. MultiWOZ 2.0 is a task-oriented              validation set. We use a batch size of 32. The
dataset of human-human written conversations               DST model is trained using ADAM (Kingma and
spanning over seven domains, consists of 10348             Ba, 2014) with the learning rate of 1e-4. During
multi-turn dialogues. MultiWOZ 2.1 is a revised            training, we use the ground truth of the previous
version of MultiWOZ 2.0, which is re-annotated             dialogue state and the ground truth value tokens.
with a different set of inter-annotators and also          In the inference, the predicted dialogue state of the
canonicalized entity names. According to the work          last turn is applied, and we use a greedy search
of Eric et al. (2019), about 32% of the state anno-        strategy in the decoding process of the value de-
tations is corrected so that the effect of noise is        coder.
counteracted.
   Note that hospital and police are excluded since        5.3     Baseline Models
they appear in training set with a very low fre-           We make a comparison with the following exist-
quency, and they do not even appear in the test set.       ing models, which are either predefined ontology-
To this end, five domains (restaurant, train, hotel,       based DSTs or open-vocabulary based DSTs. Pre-
taxi, attraction) are involved in the experiments          defined ontology-based DSTs have the advantage
with 17 distinct slots and 30 domain-slot pairs.           of access to the known candidate set of each slot,
   We follow similar data pre-processing proce-            while these approaches may not be applicable in
dures as Wu et al. (2019) on both MultiWOZ 2.0             the real scenario.
and 2.1. 2 The resulting corpus includes 8,438             FJST (Eric et al., 2019): It exploits a bidirectional
multi-turn dialogues in training set with an aver-         LSTM network to encode the dialog history and a
age of 13.5 turns per dialogue. Data statistics of         separate FFN to predict the value for each slot.
MultiWOZ 2.1 is shown in Table 1. The adjacency            HJST (Eric et al., 2019): It encodes the dialogue
matrix AG of MultiWOZ 2.0 and 2.1 datasets is              history using an LSTM like FJST, but utilizes a
shown in Figure 4 of Appendix, while domain-slot           hierarchical network.
pairs are omitted due to space limitations.                SUMBT (Lee et al., 2019): It exploits BERT (De-
                                                           vlin et al., 2018) as the encoder for the dialogue
5.2    Experiment Settings                                 context and slot-value pairs. After that, it scores
We set the hidden size of CSFN, dmodel , as 400            every candidate slot-value pair with the dialogue
with 4 heads. Following Wu et al. (2019), the              context by using a distance measure.
token embeddings with 400 dimensions are ini-              HyST (Goel et al., 2019): It is a hybrid approach
tialized by concatenating Glove embeddings (Pen-           based on hierarchical RNNs, which incorporates
nington et al., 2014) and character embeddings             both a predefined ontology-based setting and an
  2
    https://github.com/budzianowski/                       open-vocabulary setting.
multiwoz                                                   DST-Reader (Gao et al., 2019): It models the DST
Models                              BERT used     MultiWOZ 2.0     MultiWOZ 2.1
                       HJST (Eric et al., 2019)*              7              38.40            35.55
                       FJST (Eric et al., 2019)*              7              40.20            38.00
                       SUMBT (Lee et al., 2019)               3              42.40              -
                       HyST (Goel et al., 2019)*              7              42.33            38.10
          predefined
                       DS-DST (Zhang et al., 2019)            3                -              51.21
           ontology
                       DST-Picklist (Zhang et al., 2019)      3                -              53.30
                       DSTQA (Zhou and Small, 2019)           7              51.44            51.17
                       SST (Chen et al., 2020)                7              51.17            55.23
                       DST-Span (Zhang et al., 2019)          3                -              40.39
                       DST-Reader (Gao et al., 2019)*         7              39.41            36.40
                       TRADE (Wu et al., 2019)*               7              48.60            45.60
                       COMER (Ren et al., 2019)               3              48.79              -
            open-      NADST (Le et al., 2020)                7              50.52            49.04
          vocabulary   SOM-DST (Kim et al., 2019)             3              51.72            53.01
                       CSFN-DST (ours)                        7              49.59            50.81
                       CSFN-DST + BERT (ours)                 3              51.57            52.88
                       SOM-DST (our implementation)           3              51.66            52.85
                       SOM-DST + Schema Graph (ours)          3              52.23            53.19

Table 2: Joint goal accuracy (%) on the test set of MultiWOZ 2.0 and 2.1. * indicates a result borrowed from Eric
et al. (2019). 3means that a BERT model (Devlin et al., 2018) with contextualized word embeddings is utilized.

from the perspective of text reading comprehen-             jointly encode the previous state, the previous and
sions, and get start and end positions of the corre-        current dialogue utterances. An RNN-decoder is
sponding text span in the dialogue context.                 also applied to generate values for slots that need
DST-Span (Zhang et al., 2019): It treats all                to be updated in the open-vocabulary setting.
domain-slot pairs as span-based slots like DST-
Reader, and applies a BERT as the encoder.                  5.4   Main Results
DST-Picklist (Zhang et al., 2019): It defines
picklist-based slots for classification similarly to        Joint goal accuracy is the evaluation metric in our
SUMBT and applies a pre-trained BERT for the                experiments, which is represented as the ratio of
encoder. It relies on a predefined ontology.                turns whose predicted dialogue states are entirely
                                                            consistent with the ground truth in the test set.
DS-DST (Zhang et al., 2019): Similar to HyST, it
                                                            Table 2 illustrates that the joint goal accuracy of
is a hybrid system of DS-Span and DS-Picklist.
                                                            CSFN-DST and other baselines on the test set of
DSTQA (Zhou and Small, 2019): It models multi-
                                                            MultiWOZ 2.0 and MultiWOZ 2.1 datasets.
domain DST as a question answering problem, and
generates a question asking for the value of each              As shown in the table, our proposed CSFN-DST
domain-slot pair. It heavily relies on a predefined         can outperform other models except for SOM-DST.
ontology, i.e., the candidate set for each slot is          By combining our schema graphs with SOM-DST,
known, except for five time-related slots.                  we can achieve state-of-the-art performances on
                                                            both MultiWOZ 2.0 and 2.1 in the open-vocabulary
TRADE (Wu et al., 2019): It contains a slot gate
                                                            setting. Additionally, our method using BERT
module for slots classification and a pointer gener-
                                                            (Bert-base-uncased) can obtain very com-
ator for dialogue state generation.
                                                            petitive performance with the best systems in the
COMER (Ren et al., 2019): It uses a hierarchical
                                                            predefined ontology-based setting. When a BERT
decoder to generate the current dialogue state itself
                                                            is exploited, we initialize all parameters of CSFN
as the target sequence.
                                                            with the BERT encoder’s and initialize the to-
NADST (Le et al., 2020): It uses a non-                     ken/position embeddings with the BERT’s.
autoregressive decoding scheme to generate the
current dialogue state.
                                                            5.5   Analysis
SST (Chen et al., 2020): It utilizes a graph attention
matching network to fuse information from utter-            In this subsection, we will conduct some ablation
ances and schema graphs, and a recurrent graph              studies to figure out the potential factors for the im-
attention network to control state updating. How-           provement of our method. (Additional experiments
ever, it heavily relies on a predefined ontology.           and results are reported in Appendix C, case study
SOM-DST (Kim et al., 2019): It uses a BERT to               is shown in Appendix D.)
Models                                  Joint Acc. (%)      Models      Attr.   Hotel     Rest.     Taxi    Train
   CSFN-DST                                     50.81        CSFN-DST      64.78   46.29     64.64     47.35   69.79
               B
     (-) Omit HL t−1 in the decoder             48.66         (-) No SG    65.97   45.48     62.94     46.42   67.58
     (-) Omit HX t
               L in the decoder                 48.45
     (+) The previous utterance Xt−1            50.75       Table 5: Domain-specific joint accuracy on MultiWOZ
                                                            2.1. SG means Schema Graph.
Table 3: Ablation studies for context information on
MultiWOZ 2.1.                                                   Turn   Proportion (%)      w/ SG     w/o SG
                                                                  1        13.6            89.39     88.19 (−1.20)
                                             BERT used            2        13.6            73.87     72.87 (−1.00)
 Models
                                              7     3             3        13.4            58.69     57.78 (−0.91)
 CSFN-DST                                   50.81 52.88           4        12.8            51.96     50.80 (−1.16)
  (-) No schema graph, AG = 1               49.93 52.50           5        11.9            41.01     39.63 (−1.38)
  (-) No schema graph, AG = I               49.52 52.46           6        10.7            34.51     35.15 (+0.64)
  (+) Ground truth of the previous state    78.73 80.35           7         9.1            27.91     29.55 (+1.64)
  (+) Ground truth slot-gate classifi.      77.31 80.66           8         6.3            24.73     23.23 (−1.50)
  (+) Ground truth value generation         56.50 59.12           9         4.0            20.55     19.18 (−1.37)
                                                                 10         2.3            16.37     12.28 (−4.09)
                                                                 11         1.3            12.63     8.42 (−4.21)
Table 4: Joint goal accuracy(%) of ablation studies on           12         0.6            12.77     8.51 (−4.26)
MultiWOZ 2.1.                                                   > 12        0.4            9.09      0.00 (−9.09)
                                                                 all        100            50.81     49.93

5.5.1   Effect of context information                       Table 6: Joint accuracies over different dialogue turns
                                                            on MultiWOZ 2.1. It shows the impact of using schema
Context information consists of the previous dia-
                                                            graph on our proposed CSFN-DST.
logue state or the current dialogue utterance, which
are definitely key for the encoder. It would be in-
teresting to know whether the two kinds of context          slot relations is essential for joint accuracy, we
information are also essential for the RNN-based            further make analysis on domain-specific and turn-
value decoder. As shown in Table 3, we choose to            specific results. As shown in Table 5, the schema
omit the top hidden states of the previous dialogue         graph can benefit almost all domains except for
         B
state (HL t−1 ) or the current utterance (HXt
                                           L ) in the       Attaction (Attr.). As illustrated in Table 1, the At-
RNN-based value decoder. The results show both              taction domain contains only three slots, which
of them are crucial for generating real values.             should be much simpler than the other domains.
   Do we need more context? Only the current                Therefore, we may say that the schema graph can
dialogue utterance is utilized in our model, which          help complicated domains.
would be more efficient than the previous meth-                The turn-specific results are shown in Table 6,
ods involving all the preceding dialogue utterance.         where joint goal accuracies over different dialogue
However, we want to ask whether the performance             turns are calculated. From the table, we can see
will be improved when more context is used. In              that data proportion of larger turn number becomes
Table 3, it shows that incorporating the previous           smaller while the larger turn number refers to more
dialogue utterance Xt−1 gives no improvement,               challenging conversation. From the results of the
which implies that jointly encoding the current ut-         table, we can find the schema graph can make im-
terance and the previous dialogue state is effective        provements over most dialogue turns.
as well as efficient.
                                                            5.5.3   Oracle experiments
5.5.2   Effect of the schema graph                          The predicted dialogue state at the last turn is uti-
In CSFN-DST, the schema graph with domain-slot              lized in the inference stage, which is mismatched
relations is exploited. To check the effectiveness          with the training stage. An oracle experiment is
of the schema graph used, we remove knowledge-              conducted to show the impact of training-inference
aware domain-slot relations by replacing the ad-            mismatching, where ground truth of the previous
jacency matrix AG as a fully connected one 1 or             dialogue state is fed into CSFN-DST. The results
node-independent one I. Results in Table 4 show             in Table 4 show that joint accuracy can be nearly
that joint goal accuracies of models without the            80% with ground truth of the previous dialogue
schema graph are decreased similarly when BERT              state. Other oracle experiments with ground truth
is either used or not.                                      slot-gate classification and ground truth value gen-
   To reveal why the schema graph with domain-              eration are also conducted, as shown in Table 4.
5.5.4     Slot-gate classification                      value prediction that DSTQA exploits a hybrid of
We conduct experiments to evaluate our model per-       value classifier and span prediction layer, which
formance on the slot-gate classification task. Ta-      relies on a predefined ontology.
ble 7 shows F1 scores of the three slot gates, i.e.,       SOM-DST (Kim et al., 2019) is very similar to
{NONE, DONTCARE, PTR}. It seems that the pre-           our proposed CSFN-DST with BERT. The main
trained BERT model helps a lot in detecting slots       difference between SOM-DST and CSFN-DST is
of which the user doesn’t care about values. The        how to exploit the previous dialogue state. For the
F1 score of DONTCARE is much lower than the             previous dialogue state, SOM-DST considers all
others’, which implies that detecting DONTCARE          domain-slot pairs and their values (if a domain-
is a much challenging sub-task.                         slot pair contains an empty value, a special token
                                                        NONE is used), while CSFN-DST only consid-
        Gate       CSFN-DST     CSFN-DST + BERT         ers the domain-slot pairs with a non-empty value.
        NONE         99.18           99.19              Thus, SOM-DST knows which domain-slot pairs
      DONTCARE       72.50           75.96
         PTR         97.66           98.05              are empty and would like to be filled with a value.
                                                        We think that it is the strength of SOM-DST. How-
      Table 7: Slot-gate F1 scores on MultiWOZ 2.1.     ever, we choose to omit the domain-slot pairs with
                                                        an empty value for a lower computation burden,
                                                        which is proved in Table 8. As shown in the last
5.6     Reproducibility
                                                        two rows of Table 2, the schema graph can also
We run our models on GeForce GTX 2080 Ti                improve SOM-DST, which achieves 52.23% and
Graphics Cards, and the average training time for       53.19% joint accuracies on MultiWOZ 2.0 and 2.1,
each epoch and number of parameters in each             respectively. Appendix E shows how to exploit
model are provided in Table 8. If BERT is ex-           schema graph in SOM-DST.
ploited, we accumulate the gradients with 4 steps
for a minibatch of data samples (i.e., 32/4 = 8         6   Conclusion and Future Work
samples for each step), due to the limitation of
GPU memory. As mentioned in Section 5.4, joint          We introduce a multi-domain dialogue state tracker
goal accuracy is the evaluation metric used in our      with context and schema fusion networks, which
experiments, and we follow the computing script         involves slot relations and learns deep representa-
provided in TRADE-DST 3 .                               tions for each domain-slot pair dependently. Slots
                                                        from different domains and their relations are or-
     Method             Time per Batch   # Parameters   ganized as a schema graph. Our approach outper-
    CSFN-DST                350ms            63M
 CSFN-DST + BERT            840ms            115M       forms strong baselines on both MultiWOZ 2.0 and
  SOM-DST + SG             1160ms            115M       2.1 benchmarks. Ablation studies also show that
                                                        the effectiveness of the schema graph.
  Table 8: Runtime and mode size of our methods.
                                                           It will be a future work to incorporate rela-
                                                        tions among dialogue states, utterances and domain
5.7     Discussion                                      schemata. To further mitigate the data sparsity
                                                        problem of multi-domain DST, it would be also in-
The main contributions of this work may focus on
                                                        teresting to incorporate data augmentations (Zhao
exploiting the schema graph with graph-based at-
                                                        et al., 2019) and semi-supervised learnings (Lan
tention networks. Slot-relations are also utilized
                                                        et al., 2018; Cao et al., 2019).
in DSTQA (Zhou and Small, 2019). However,
DSTQA uses a dynamically-evolving knowledge
                                                        Acknowledgments
graph for the dialogue context, and we use a static
schema graph. We absorb the dialogue context            We thank the anonymous reviewers for their
by using the previous (predicted) dialogue state        thoughtful comments.     This work has been
as another input. We believe that the two different     supported by Shanghai Jiao Tong University
usages of the slot relation graph can be complemen-     Scientific and Technological Innovation Funds
tary. Moreover, these two methods are different in      (YG2020YQ01) and No. SKLMCPTS2020003
  3
    https://github.com/jasonwu0731/                     Project.
trade-dst
References                                                 Rahul Goel, Shachi Paul, and Dilek Hakkani-Tür.
                                                             2019. HyST: A hybrid approach for flexible and
Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hin-          accurate dialogue state tracking. arXiv preprint
   ton. 2016. Layer normalization. arXiv preprint            arXiv:1907.00883.
   arXiv:1607.06450.
                                                           Kazuma Hashimoto, Caiming Xiong, Yoshimasa Tsu-
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben-             ruoka, and Richard Socher. 2017. A joint many-task
  gio. 2014. Neural machine translation by jointly           model: Growing a neural network for multiple NLP
  learning to align and translate. arXiv preprint            tasks. In Proceedings of the 2017 Conference on
  arXiv:1409.0473.                                           Empirical Methods in Natural Language Processing,
                                                             pages 1923–1933.
Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang
  Tseng, Iñigo Casanueva, Stefan Ultes, Osman Ra-         Matthew Henderson, Blaise Thomson, and Jason D
  madan, and Milica Gašić. 2018. MultiWOZ - a             Williams. 2014a. The second dialog state tracking
  large-scale multi-domain wizard-of-Oz dataset for         challenge. In Proceedings of Annual SIGdial Meet-
  task-oriented dialogue modelling. In Proceedings of       ing on Discourse and Dialogue, pages 263–272.
  the 2018 Conference on Empirical Methods in Natu-
  ral Language Processing, pages 5016–5026.
                                                           Matthew Henderson, Blaise Thomson, and Steve
                                                            Young. 2014b. Robust dialog state tracking using
Ruisheng Cao, Su Zhu, Chen Liu, Jieyu Li, and Kai
                                                            delexicalised recurrent neural networks and unsuper-
  Yu. 2019. Semantic parsing with dual learning. In
                                                            vised adaptation. In Proceedings of IEEE Spoken
  Proceedings of the 57th Annual Meeting of the Asso-
                                                            Language Technology Workshop (SLT).
  ciation for Computational Linguistics, pages 51–64.
                                                           Matthew Henderson, Blaise Thomson, and Steve
Lu Chen, Zhi Chen, Bowen Tan, Sishan Long, Mil-
                                                            Young. 2014c. Word-based dialog state tracking
  ica Gasic, and Kai Yu. 2019. AgentGraph: To-
                                                            with recurrent neural networks. In Proceedings of
  wards universal dialogue management with struc-
                                                            Annual SIGdial Meeting on Discourse and Dialogue,
  tured deep reinforcement learning. IEEE/ACM
                                                            pages 292–299.
  Transactions on Audio, Speech, and Language Pro-
  cessing, 27(9):1378–1391.
                                                           Mettew Henderson, Blaise Thomson, and Jason D.
Lu Chen, Boer Lv, Chi Wang, Su Zhu, Bowen Tan,              Williams. 2014d. The third dialog state tracking
  and Kai Yu. 2020. Schema-guided multi-domain di-          challenge. In Proceedings of IEEE Spoken Lan-
  alogue state tracking with graph attention neural net-    guage Technology Workshop (SLT).
  works. In AAAI, pages 7521–7528.
                                                           Sungdong Kim, Sohee Yang, Gyuwan Kim, and Sang-
Lu Chen, Bowen Tan, Sishan Long, and Kai Yu. 2018.           Woo Lee. 2019. Efficient dialogue state tracking
  Structured dialogue policy with graph neural net-          by selectively overwriting memory. arXiv preprint
  works. In Proceedings of the 27th International            arXiv:1911.03906.
  Conference on Computational Linguistics, pages
  1257–1268.                                               Diederik P Kingma and Jimmy Ba. 2014. Adam: A
                                                             method for stochastic optimization. arXiv preprint
Kyunghyun Cho, Bart Van Merriënboer, Dzmitry Bah-           arXiv:1412.6980.
  danau, and Yoshua Bengio. 2014. On the properties
  of neural machine translation: Encoder-decoder ap-       Ouyu Lan, Su Zhu, and Kai Yu. 2018. Semi-supervised
  proaches. arXiv preprint arXiv:1409.1259.                  training using adversarial multi-task learning for
                                                             spoken language understanding. In 2018 IEEE Inter-
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and                national Conference on Acoustics, Speech and Sig-
   Kristina Toutanova. 2018. BERT: Pre-training of           nal Processing (ICASSP), pages 6049–6053. IEEE.
   deep bidirectional transformers for language under-
   standing. arXiv preprint arXiv:1810.04805.              Hung Le, Richard Socher, and Steven C.H. Hoi. 2020.
                                                             Non-autoregressive dialog state tracking. In Interna-
Mihail Eric, Rahul Goel, Shachi Paul, Abhishek Sethi,        tional Conference on Learning Representations.
  Sanchit Agarwal, Shuyag Gao, and Dilek Hakkani-
  Tur. 2019. MultiWOZ 2.1: Multi-domain dialogue           Hwaran Lee, Jinsik Lee, and Tae-Yoon Kim. 2019.
  state corrections and state tracking baselines. arXiv      SUMBT: Slot-utterance matching for universal and
  preprint arXiv:1907.01669.                                 scalable belief tracking. In Proceedings of the 57th
                                                            Annual Meeting of the Association for Computa-
Shuyang Gao, Abhishek Sethi, Sanchit Agarwal, Tagy-          tional Linguistics, pages 5478–5483.
  oung Chung, and Dilek Hakkani-Tur. 2019. Dia-
  log state tracking: A neural reading comprehension       Diego Marcheggiani, Joost Bastings, and Ivan Titov.
  approach. In Proceedings of the 20th Annual SIG-           2018. Exploiting semantics in neural machine trans-
  dial Meeting on Discourse and Dialogue, pages 264–         lation with graph convolutional networks. arXiv
  273.                                                       preprint arXiv:1804.08313.
Nikola Mrkšić, Diarmuid Ó Séaghdha, Tsung-Hsien        Jason Williams, Antoine Raux, Deepak Ramachandran,
  Wen, Blaise Thomson, and Steve Young. 2017. Neu-            and Alan Black. 2013. The dialog state tracking
  ral belief tracker: Data-driven dialogue state track-       challenge. In Proceedings of Annual SIGdial Meet-
  ing. In Proceedings of the 55th Annual Meeting              ing on Discourse and Dialogue, pages 404–413.
  of the Association for Computational Linguistics,
  pages 1777–1788.                                         Jason D Williams. 2012. A belief tracking challenge
                                                              task for spoken dialog systems. In NAACL-HLT
Jeffrey Pennington, Richard Socher, and Christopher          Workshop on Future Directions and Needs in the
   Manning. 2014. Glove: Global vectors for word rep-         Spoken Dialog Community: Tools and Data, pages
   resentation. In Proceedings of the 2014 Conference         23–24.
   on Empirical Methods in Natural Language Process-
   ing (EMNLP), pages 1532–1543.                           Chien-Sheng Wu, Andrea Madotto, Ehsan Hosseini-
                                                             Asl, Caiming Xiong, Richard Socher, and Pascale
Osman Ramadan, Paweł Budzianowski, and Milica Ga-            Fung. 2019. Transferable multi-domain state gener-
  sic. 2018. Large-scale multi-domain belief track-          ator for task-oriented dialogue systems. In Proceed-
  ing with knowledge sharing. In Proceedings of the          ings of the 57th Annual Meeting of the Association
  56th Annual Meeting of the Association for Compu-          for Computational Linguistics.
  tational Linguistics, pages 432–437.

Liliang Ren, Jianmo Ni, and Julian McAuley. 2019.          Puyang Xu and Qi Hu. 2018. An end-to-end approach
   Scalable and accurate dialogue state tracking via hi-     for handling unknown slot values in dialogue state
   erarchical sequence generation. In Proceedings of         tracking. In Proceedings of the 56th Annual Meeting
   the 2019 Conference on Empirical Methods in Nat-          of the Association for Computational Linguistics.
   ural Language Processing and the 9th International
   Joint Conference on Natural Language Processing         Liang Yao, Chengsheng Mao, and Yuan Luo. 2019.
   (EMNLP-IJCNLP), pages 1876–1885.                          Graph convolutional networks for text classification.
                                                             In Proceedings of the AAAI Conference on Artificial
Liliang Ren, Kaige Xie, Lu Chen, and Kai Yu. 2018.           Intelligence, pages 7370–7377.
   Towards universal dialogue state tracking. In Pro-
   ceedings of the 2018 Conferenceon Empirical Meth-       Steve Young, Milica Gašić, Blaise Thomson, and Ja-
   ods in Natural Language Processing (EMNLP),                son D Williams. 2013. Pomdp-based statistical spo-
   pages 2780–2786.                                           ken dialog systems: A review. Proceedings of the
                                                              IEEE, 101(5):1160–1179.
Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus
  Hagenbuchner, and Gabriele Monfardini. 2009. The         Kai Yu, Kai Sun, Lu Chen, and Su Zhu. 2015.
  graph neural network model. IEEE Transactions on           Constrained markov bayesian polynomial for effi-
  Neural Networks, 20(1):61–80.                              cient dialogue state tracking. IEEE/ACM Transac-
                                                             tions on Audio, Speech, and Language Processing,
Abigail See, Peter J Liu, and Christopher D Man-             23(12):2177–2188.
  ning. 2017. Get to the point: Summarization
  with pointer-generator networks. arXiv preprint          Jian-Guo Zhang, Kazuma Hashimoto, Chien-Sheng
  arXiv:1704.04368.                                           Wu, Yao Wan, Philip S Yu, Richard Socher, and
                                                              Caiming Xiong. 2019. Find or classify? dual strat-
Kai Sun, Lu Chen, Su Zhu, and Kai Yu. 2014a. A gen-           egy for slot-value predictions on multi-domain dia-
  eralized rule based tracker for dialogue state track-       log state tracking. arXiv preprint arXiv:1910.03544.
  ing. In Proceedings of IEEE Spoken Language Tech-
  nology Workshop (SLT).
                                                           Zijian Zhao, Su Zhu, and Kai Yu. 2019. Data augmen-
Kai Sun, Lu Chen, Su Zhu, and Kai Yu. 2014b. The              tation with atomic templates for spoken language
  SJTU system for dialog state tracking challenge 2.          understanding. In Proceedings of the 2019 Confer-
  In Proceedings of the 15th Annual Meeting of the            ence on Empirical Methods in Natural Language
  Special Interest Group on Discourse and Dialogue            Processing and the 9th International Joint Confer-
  (SIGDIAL), pages 318–326.                                   ence on Natural Language Processing (EMNLP-
                                                              IJCNLP), pages 3628–3634.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
  Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz            Victor Zhong, Caiming Xiong, and Richard Socher.
  Kaiser, and Illia Polosukhin. 2017. Attention is all       2018. Global-locally self-attentive encoder for di-
  you need. In Advances in neural information pro-           alogue state tracking. In Proceedings of the 56th An-
  cessing systems.                                           nual Meeting of the Association for Computational
                                                             Linguistics, pages 1458–1467.
Petar Veličković, Guillem Cucurull, Arantxa Casanova,
  Adriana Romero, Pietro Liò, and Yoshua Bengio.          Li Zhou and Kevin Small. 2019. Multi-domain dia-
  2018. Graph attention networks. In Proceedings              logue state tracking as dynamic knowledge graph
  of International Conference on Learning Represen-           enhanced question answering.      arXiv preprint
  tations.                                                    arXiv:1911.06192.
A        Context- and Schema-Aware Encoding                        GRU is initialized with
                                                                                                        B
Besides the hidden states HG
                           i of the schema graph
                                                                                      0
                                                                                     gtj = HX t     t−1
                                                                                            L,0 + HL,0
                                          Bt−1
G, we show the details of updating HX  t
                                     i , Hi    in
the i-th layer of CSFN:                                            and e0tj = HG L,M +N +j .
                                                                      The value generator transforms the hidden state
                 B                                     Bt−1
    HG      Xt       t−1              G    Xt
     i+1 , Hi+1 , Hi+1 = CSFNLayeri (Hi , Hi , Hi             )    to the probability distribution over the token vo-
                                                                   cabulary at the k-th step, which consists of two
                                                                   parts: 1) distribution over all input tokens, 2) distri-
   The hidden states of the dialogue utterance Xt                  bution over the input vocabulary. The first part is
at the i-the layer are updated as follows:                         computed as
                                                                                                                  B
        IXX = MultiHeadΘXX (HX t  Xt                                 Ptjctx,k = softmax(ATT(gtj
                                                                                             k
                                                                                                , [HXt
                                                                                                    L ; HL
                                                                                                           t−1
                                                                                                               ]))
                             i , Hi )
                                    Bt−1
    EXB = MultiHeadΘXB (HX t
                         i , Hi             )                      where Ptjctx,k ∈ R1×(|Xt |+|Bt−1 |) , and ATT(., .) is a
    EXG = MultiHeadΘXG (HX t  G                                    function to get attention weights (Bahdanau et al.,
                         i , Hi )
                                                                   2014) with more details shown in Appendix B.1.
        CX = LayerNorm(HXt
                        i + IXX + EXB + EXG )                      The second part is calculated as
    HX t
     i+1 = LayerNorm(CX + FFN(CX ))                                                                      B
                                                                                 cktj = Ptjctx,k [HXt
                                                                                                   L ; HL
                                                                                                          t−1
                                                                                                              ]
where IXX contains internal attention vectors, EXB                         Ptjvocab,k = softmax([gtj
                                                                                                  k
                                                                                                     ; cktj ]Wproj E > )
and EXG are external attention vectors.
  The hidden states of the previous dialogue state                 where Ptjvocab,k ∈ R1×dvocab , cktj ∈ R1×dmodel is a
Bt−1 at the i-the layer are updated as follows:                    context vector, Wproj ∈ R2dmodel ×dmodel is a trainable
                                   Bt−1         Bt−1
                                                                   parameter, and E ∈ Rdvocab ×dmodel is the token em-
     IBB = GraphMultiHeadΘBB (Hi          , Hi         , ABt−1 )   bedding matrix shared across the encoder and the
                            Bt−1                                   decoder.
    EBX = MultiHeadΘBX (Hi         , HX t
                                      i )
                            Bt−1                                     We use the soft copy mechanism (See et al.,
    EBG = MultiHeadΘBG (Hi         , HG
                                      i )                          2017) to get the final output distribution over all
                         Bt−1
    CB = LayerNorm(Hi           + IBB + EBX + EBG )                candidate tokens:
    B
                                                                         Ptjvalue,k = pgen Ptjvocab,k + (1 − pgen )Ptjctx,k
   t−1
Hi+1   = LayerNorm(CB + FFN(CB ))
                                                                                              k
                                                                             pgen = sigmoid([gtj ; ektj ; cktj ]Wgen )
where ABt−1 is the adjacency matrix of the previ-
ous dialogue state. The adjacency matrix indicates                 where Wgen ∈ R3dmodel ×1 is a trainable parameter.
that each triplets in Bt−1 is separated, while tokens                The loss function for value decoder is
in a same triplet are connected with each other. The                              T X X
[CLS] token is connected with all triplets, serving
                                                                                  X
                                                                   Lvalue = −                    log(Ptjvalue,k · (ytj
                                                                                                                    value,k >
                                                                                                                           ) )
as a transit node.                                                                t=1 j∈Ct k

B       RNN-based Value Decoder                                           value,k
                                                                   where ytj      is the one-hot token label for the j-th
                                                                   domain-slot pair at k-th step.
After the slot-gate classification, there are J 0
domain-slot pairs tagged with PTR class which                      B.1     Attention Weights
indicates the domain-slot should take a real value.                For attention mechanism for computing Ptjctx,k in
                                           gate
They are denoted as Ct = {j|argmax(Ptj ) =                         the RNN-based value decoder, we follow Bahdanau
             0
PTR}, and J = |Ct |.                                               et al. (2014) and define the ATT(., .) function as
   We use Gated Recurrent Unit (GRU) (Cho et al.,
2014) decoder like Wu et al. (2019) and See et al.
                                                                                                      att >
                           k ∈ R1×dmodel is recur-
(2017). The hidden state gtj                                                ui =tanh(xWatt       att
                                                                                         1 + hi W2 + b )v
sively updated by taking a word embedding ektj as                                   exp(ui )
                                                                            ai = PS
the input until [EOS] token is generated:
                                                                                   j=1 exp(uj )
                  k
                 gtj        k−1 k
                     = GRU(gtj , etj )                                       a ={a1 , · · · , aS } = ATT(x, H)
where x ∈ R1×d , H ∈ RS×d , Watt 1 ∈R
                                      d×d , Watt ∈
                                             2          where Dt−1 and Dt are the last and current ut-
  d×d    att     1×d         1×d
R , b ∈ R , v ∈ R , and hi is the i-th                  terances, respectively. The dialogue state Bt is
row vector of H. Therefore, ATT(x, H) returns an        denoted as Bt = Bt1 ⊕ . . . ⊕ BtJ , where Btj =
attention distribution of x over H.                     [SLOT]j ⊕ S j ⊕ − ⊕ Vtj is the representation of
                                                        the j-th slot-value pair. To incorporate the schema
C    Additional Results                                 graph, we exploit the special token [SLOT]j to re-
Domain-specific Results Domain-specific accu-           place the domain-slot node oj in the schema graph
racy is the joint goal accuracy measured on a subset    (j = 1, · · · , J). Then, domain and slot nodes
of the predicted dialogue state, which only contains    G0 = d1 ⊕ · · · ⊕ dM ⊕ s1 ⊕ · · · ⊕ sN are con-
the slots belong to a domain. From the results of       catenated into Xt , i.e.,
Table 9, we can find BERT can make improvements
                                                            Xt = [CLS] ⊕ Dt−1 ⊕ Dt ⊕ Bt−1 ⊕ G0 ,
on all domains, and especially the improvement on
Taxi domain is the largest.                             where the relations among domain, slot and
     Domain      CSFN-DST     CSFN-DST + BERT           domain-slot nodes are also considered in attention
    Attraction     64.78           67.82                masks of BERT.
      Hotel        46.29           48.80
    Restaurant     64.64           65.23
       Taxi        47.35           53.58
      Train        69.79           70.91

Table 9: Domain-specific joint accuracy on MultiWOZ
2.1.

Slot-specific Results Slot-specific F1 score is mea-
sured for predicting slot-value pairs of the corre-
sponding slot. Table 10 shows slot-specific F1
scores of CSFN-DST without the schema graph,
CSFN-DST and CSFN-DST with BERT on the test
set of MultiWOZ 2.1.

D    Case Study
We also conduct case study on the test set of Mul-
tiWOZ 2.1, and four cases are shown in Table 11.
From the first three cases, we can see the schema
graph can copy values from related slots in the
memory (i.e., the previous dialogue state). In the
case C1, the model makes the accurate reference
of the phrase “whole group” through the context,
and the value of restaurant-book people is copied
as the value of train-book people. We can also
see a failed case (C4). It is too complicated to in-
ference the departure and destination by a word
“commute”.

E    SOM-DST with Schema Graph
For SOM-DST (Kim et al., 2019), the input tokens
to the state operation predictor are the concatena-
tion of the previous turn dialog utterances, the cur-
rent turn dialog utterances, and the previous turn
dialog state:

       Xt = [CLS] ⊕ Dt−1 ⊕ Dt ⊕ Bt−1 ,
Figure 4: Adjacency matrix AG of MultiWOZ 2.0 and 2.1 datasets. It contains only domain and slot nodes,
while domain-slot paris are omitted due to space limitations. The first five items are domains (“attraction, hotel,
restaurant, taxi, train”), and the rest are slots.
Domain-slot        CSFN-DST (no SG)       CSFN-DST       CSFN-DST + BERT
                attraction-area           91.67               91.92             92.81
               attraction-name            78.28               77.77             79.55
                attraction-type           90.95               90.89             91.97
                     hotel-area           84.59               84.21             84.86
               hotel-book day             97.16               97.79             97.03
             hotel-book people            95.43               96.14             97.35
               hotel-book stay            96.04               97.00             96.98
                 hotel-internet           86.79               89.98             86.58
                    hotel-name            82.97               83.11             84.61
                 hotel-parking            86.68               87.66             86.07
              hotel-price range           89.09               90.10             92.56
                    hotel-stars           91.51               93.49             93.34
                    hotel-type            77.58               77.87             82.12
                restaurant-area           93.73               94.27             94.09
            restaurant-book day           97.75               97.66             97.42
         restaurant-book people           96.79               96.67             97.84
           restaurant-book time           92.43               91.60             94.29
               restaurant-food            94.90               94.18             94.48
              restaurant-name             80.72               81.39             80.59
          restaurant-price range          93.47               94.28             94.49
                 taxi-arrive by           78.81               81.09             86.08
                 taxi-departure           73.15               71.39             75.15
               taxi-destination           73.79               78.06             79.83
                   taxi-leave at          77.29               80.13             88.06
                train-arrive by           86.43               87.56             88.77
             train-book people            89.59               91.41             92.33
                     train-day            98.44               98.41             98.52
                train-departure           95.91               96.52             96.22
              train-destination           97.08               97.06             96.13
                  train-leave at          69.97               70.97             74.50
             Joint Acc. overall           49.93               50.81             52.88

Table 10: Slot-specific F1 scores on MultiWOZ 2.1. SG means Schema Graph. The results in bold black are the
best slot F1 scores.
(restaurant-book day, friday), (restaurant-book people, 8), (restaurant-book time, 10:15),
      Previous DS:          (restaurant-name, restaurant 2 two), (train-leave at, 12:15),
                            (train-destination, peterborough), (train-day, saturday), (train-departure, cambridge)
      System:               How about train tr3934? It leaves at 12:34 & arrives at 13:24. Travel time is 50 minutes.
      Human:                That sounds fine. Can I get tickets for my whole group please?
                            (restaurant-name, restaurant 2 two), (restaurant-book day, friday),
                            (restaurant-book people, 8), (restaurant-book time, 10:15), (train-departure, cambridge),
      Gold DS:
                            (train-leave at, 12:15), (train-day, saturday), (train-destination, peterborough),
                            (train-book people, 8)
 C1                         (restaurant-name, restaurant 2 two), (restaurant-book day, friday),
                            (restaurant-book people, 8), (restaurant-book time, 10:15), (train-departure, cambridge),
      CSFN-DST (no SG):
                            (train-leave at, 12:15), (train-day, saturday), (train-destination, peterborough),
                            (train-book people, 1)
                            (restaurant-name, restaurant 2 two), (restaurant-book day, friday),
                            (restaurant-book people, 8), (restaurant-book time, 10:15), (train-departure, cambridge),
      CSFN-DST:
                            (train-leave at, 12:15), (train-day, saturday), (train-destination, peterborough),
                            (train-book people, 8)
                            (hotel-area, west), (hotel-price range, cheap), (hotel-type, guest house),
      Previous DS:          (hotel-internet, yes), (hotel-name, warkworth house), (restaurant-area, centre),
                            (restaurant-food, italian), (restaurant-price range, cheap), (restaurant-name, ask)
      System:               01223364917 is the phone number. 12 bridge street city centre, cb21uf is the address.
      Human:                Thanks. I will also need a taxi from the hotel to the restaurant. Will you handle this?
                            (hotel-area, west), (hotel-price range, cheap), (hotel-type, guest house),
                            (hotel-internet, yes), (hotel-name, warkworth house), (restaurant-area, centre),
      Gold DS:
                            (restaurant-food, italian), (restaurant-price range: cheap), (restaurant-name, ask),
                            (taxi-departure, warkworth house), (taxi-destination, ask)
 C2                         (hotel-area, west), (hotel-price range, cheap), (hotel-type, guest house),
                            (hotel-internet, yes), (hotel-name, warkworth house), (restaurant-area, centre),
      CSFN-DST (no SG):
                            (restaurant-food, italian), (restaurant-price range: cheap), (restaurant-name, ask),
                            (taxi-departure, warkworth house), (taxi-destination, warkworth house)
                            (hotel-area, west), (hotel-price range, cheap), (hotel-type, guest house),
                            (hotel-internet, yes), (hotel-name, warkworth house), (restaurant-area, centre),
      CSFN-DST:
                            (restaurant-food, italian), (restaurant-price range: cheap), (restaurant-name, ask),
                            (taxi-departure, warkworth house), (taxi-destination, ask)
                            (attraction-area, east), (attraction-name, funky fun house), (restaurant-area, east),
      Previous DS:          (restaurant-food, indian), (restaurant-price range, moderate),
                            (restaurant-name, curry prince)
      System:               cb58jj is there postcode. Their address is 451 newmarket road fen ditton.
                            Great, thank you! Also, can you please book me a taxi between the restaurant and funky
      Human:
                            fun house? I want to leave the restaurant by 01:30.
                            (attraction-area, east), (attraction-name, funky fun house), (restaurant-area, east),
                            (restaurant-food, indian), (restaurant-price range, moderate),
      Gold DS:
                            (restaurant-name, curry prince), (taxi-departure, curry prince),
                            (taxi-destination, funky fun house), (taxi-leave at, 01:30)
 C3
                            (attraction-area, east), (attraction-name, funky fun house), (restaurant-area, east),
                            (restaurant-food, indian), (restaurant-price range, moderate),
      CSFN-DST (no SG):
                            (restaurant-name, curry prince), (taxi-departure, curry garden),
                            (taxi-destination, funky fun house), (taxi-leave at, 01:30)
                            (attraction-area, east), (attraction-name, funky fun house), (restaurant-area, east),
                            (restaurant-food, indian), (restaurant-price range, moderate),
      CSFN-DST:
                            (restaurant-name, curry prince), (taxi-departure, curry prince),
                            (taxi-destination, funky fun house), (taxi-leave at, 01:30)
                            (hotel-name, a and b guest house), (hotel-book day, tuesday), (hotel-book people, 6),
      Previous DS:
                            (hotel-book stay, 4), (attraction-area, west), (attraction-type, museum)
                            Cafe jello gallery has a free entrance fee. The address is cafe jello gallery, 13 magdalene
      System:
                            street and the post code is cb30af. Can I help you with anything else?
      Human:                Yes please. I need a taxi to commute.
                            (hotel-name, a and b guest house), (hotel-book day, tuesday), (hotel-book people, 6),
      Gold DS:              (hotel-book stay, 4), (attraction-area, west), (attraction-type, museum),
 C4
                            (taxi-destination, cafe jello gallery), (taxi-departure, a and b guest house)
                            (hotel-name, a and b guest house), (hotel-book day, tuesday), (hotel-book people, 6),
      CSFN-DST (no SG):
                            (hotel-book stay, 4), (attraction-area, west), (attraction-type, museum)
                            (hotel-name, a and b guest house), (hotel-book day, tuesday), (hotel-book people, 6),
      CSFN-DST:             (hotel-book stay, 4), (attraction-area, west), (attraction-type, museum),
                            (taxi-destination, cafe jello gallery)

Table 11: Four cases on the test set of MultiWOZ 2.1. DS means Dialogue State, and SG means Schema Graph.
You can also read